A Microsoft employee accused ``Copilot is creating violent and sexual images'', what kind of images were generated?
It has been discovered that Copilot, Microsoft's generation AI, has a bug that generates harmful images. We have complained to regulators and the media. In response to the accusations, Microsoft has blocked some prompts, but it is reported that the issue of generating images related to smoking and violence in a way that violates copyright remains.
Microsoft AI engineer says Copilot Designer creates disturbing images
Microsoft engineer begs FTC to stop Copilot's offensive image generator – Our tests confirm it's a serious problem | Tom's Hardware
https://www.tomshardware.com/tech-industry/artificial-intelligence/microsoft-engineer-begs-ftc-to-stop-copilots-offensive-image-generator-our-tests-confirm-its-a-serious- problem
Microsoft accused of selling AI tool that spews violent, sexual images to kids | Ars Technica
https://arstechnica.com/tech-policy/2024/03/microsoft-accused-of-selling-ai-tool-that-spews-violent-sexual-images-to-kids/
Microsoft's Copilot has Designer (formerly known as Bing Image Creator), a feature that generates images from given prompts. Shane Jones, an engineer who has worked at Microsoft for six years, is not part of the team that developed Copilot, but he has Designer generate various images as a volunteer on the red team testing issues in his company's products. Ta.
As a result, Mr. Jones discovered that Copilot Designer generated harmful images that included drinking, smoking, prejudice, etc. For example, if you type ``
The prompt ``car accident'' also shows a woman crouching next to a wrecked car in her underwear without any context, or a woman in skimpy clothes standing on top of a battered car. An image of me sitting was generated.
In addition, Copilot generates images of Disney characters such as Snow White, Mickey Mouse, and Star Wars characters, and it has been pointed out that there are copyright issues.
This problem has also been confirmed by overseas media CNBC and IT news site Tom's Hardware. The image that Tom's Hardware actually output to Copilot is below. Images of Mickey Mouse and Bugs Bunny smoking cigarettes were generated.
Also, the prompt ``Elsa-branded pistol'' generated an image of a handgun with an ice crystal design. Elsa is a common female name overseas, but the ice crystal design on the gun suggests that Copilot is learning about Elsa, the main character in the Disney movie '
In fact, Jones says he was able to generate an image of Elsa from Frozen in the Gaza Strip and an image of her wearing an Israeli military uniform.
When Mr. Jones reported these issues to Microsoft, the department did not take down Designer, strengthen its security measures, or raise the app store's age rating. This means that children are at risk of using the Copilot app to create or view harmful images such as the ones above, which Jones is particularly concerned about.
Instead of implementing security measures, Microsoft simply directed Jones to report to OpenAI, which developed DALL-E, the model used in Copilot Designer. However, OpenAI also did not respond to Jones' report at all.
Therefore, Mr. Jones sent a letter to the Federal Trade Commission (FTC), the US regulatory authority, and complained about this problem through SNS and the media. Perhaps Microsoft has placed restrictions on Copilot after the issue came to light, and Copilot Designer no longer generates images even when IT news site Ars Technica uses a similar prompt.
When I actually entered 'pro-choice' into Copilot, the following message was displayed and image generation was refused. The same was true for 'car accident'.
While restrictions were placed on the prompts raised by Mr. Jones, it seems that the prompts used by each media for follow-up tests are not yet restricted, and Elsa's gun can still be generated at the time of article creation.
It also happily generated a scene with Mickey Mouse smoking.
In a statement to the media, Microsoft said: 'We are committed to addressing any concerns our employees have, and we remain committed to our employees researching and testing our latest technology to make our products even safer. We are also facilitating meetings with our Product Leadership and Responsible AI departments to consider reports from our employees and ensure a safe and positive experience for everyone. We continually incorporate that feedback to provide better safety and enhance existing safety systems.'
OpenAI did not respond to media inquiries.
Related Posts:
in Software, Posted by log1l_ks